VASparse introduces a novel decoding algorithm that makes Large Vision-Language Models more trustworthy by reducing hallucinations without sacrificing speed.

• Reduces visual hallucinations by selectively filtering tokens using visual-aware sparsification
• Maintains inference efficiency unlike methods requiring secondary decoding
• Achieves 20.6% reduction in hallucination rate while improving response quality
• Works as a plug-and-play solution compatible with various LVLMs

Security Impact: By ensuring AI outputs remain faithful to visual reality, VASparse addresses a critical security concern in AI systems where hallucinated content could lead to misinformation or unreliable automated decision-making.

VASparse: Towards Efficient Visual Hallucination Mitigation via Visual-Aware Token Sparsification